Comment by bick_nyers

Just to add onto this point, you expect different experts to be activated for every token, so not having all of the weights in fast memory can still be quite slow as you need to load/unload memory every token.